[elephant-devel] Query System

[Ian Eslick]

On May 10, 2008, at 12:26 AM, lists at infoway.net wrote:

> Hi Ian,
>
> Thanks for your comments. They do make some points a bit clearer and  
> bring others to the table. I'd like to see others comment as well  
> before we continue moving forward.
>
> In summary, I agree to follow your suggestion. However, one thing  
> that still remains unclear to me is the type of result expected from  
> the query system.

See my comments below, but I think we want to avoid consing large sets  
where possible.  So once we've defined a set of objects using a query  
expressions we may want to:

1) Perform a non-consing map operation over it
2) Return it in a list
3) intersect it with another query set
4) returning a list of slot values or mapping over the slot value

#2 is easy to derive from #1 - that's what get-instances-by-value does  
today.

#3, to be efficient, may require lazy query evaluation.

#4 is what you discuss below.

I think the general idea is that a query defines a set of objects, but  
how you use it defines the time/space cost of evaluating that set.  I  
actually like the Django notion of lazy query sets - you can perform  
operations like intersecting two query sets, but the actual query  
isn't executed until you need to operate on a member of the set.  This  
gives the query compiler the maximum information to optimize.

> I was under the assumption that we want to continue handling objects  
> of the persistent classes in Elephant. Thus, the results from the  
> query should be a set (list, etc) of objects matching the criteria.  
> However, reading your comments makes me think that what you're  
> suggesting is that the result set be something closer to what a SQL  
> query would return. So, going back to my example, instead of  
> receiving a set of Books, you would get back something like:

Wow, that was not the impression I intended to make!  I definitely  
prefer an object-centric view by default.  We may want to return pairs  
of objects, however.  Returning values is easy to implement on top of  
a map, as the example below illustrates.

> ((book_oid book_title book_author publisher_oid publisher_name  
> publisher_year) (value11 value12 value 13, value14, value15 value16)  
> (value21 value22 value23 value24 value25 value26) ... )
>
> In other words, maybe something like a list that contains a list of  
> the slot names returned and then the list of matching values, where  
> the list of slots is either the concatenation of all the slots in  
> the persistent classes (in the case of something like SELECT  
> Books.*, Publishers.* ...) or the specific slots requested (e.g.  
> SELECT Books.author, Publishers.name ...).

Of course it's easy to do the above with a map operation or wrapper  
with a closure that binds map-fn, the user's function, and slots, a  
list of slots to operate on.

(lambda (obj)
   (apply #'map-fn
  	(mapcar (lambda (slot) (funcall accessor obj))
	        slots)))

So the user would see something like:

(map-query (:slots name city state country)
    (lambda (name city state country)
       ...))

It would mostly be a convenience function built on top of the basic  
query mechanism, and not essential to start with...


> And therefore, the result set would be treated as simply a set of  
> values instead of a set of objects. Within the result set, you could  
> include the OIDs so that you could eventually instantiate the  
> particular object and work in via the object model, or you could  
> simply just use the values returned, which is what you originally  
> asked for in the SELECT statement.

I do think the semantics of a query should be that query results are a  
set of objects, rather than only persistent objects.  It means the  
user can work with a result list, or map over the results; working  
with whatever my query string specifies.  That said, we should start  
by just extracting sets of objects for users to map over that satisfy  
the constraints.  We definitely should be able to specify that the  
system only return oids.

> If my assumption is correct from what you're saying, that certainly  
> clarifies a lot or my concerns and doubts. I'm sure more details  
> will arise along the way, but that could be a starting point to  
> sketch the system.
>
> Comments?
>
> Thanks,
> Daniel
>
> On May 9, 2008, at 9:54 PM, Ian Eslick wrote:
>
>> Welcome back Daniel, we all know the work drill!
>>
>> Here are a few thoughts to throw into the mix...
>>
>> One advantage of the relational model is that you have implicit  
>> data structures (tables) that can be assembled from existing tables  
>> via the SQL query.  This is nice because it means we don't have to  
>> explicitly create and maintain the structure for all these derived  
>> data structures.  In a pure lisp model, you actually have to do all  
>> this maintenance yourself, especially the optimizations necessary  
>> for efficiency that add to complexity.  I feel that Elephant should  
>> probably fall somewhere in-between.  You maintain the data  
>> structures that you want to work with in your program logic, but  
>> the system can maintain pointers and indices and other  
>> relationships that make it easy and efficient to generate and work  
>> with subsets of objects efficiently (a user's inbox, for example).
>>
>> Some of the limitations/frustrations with the current system may be  
>> caused by people trying to do familiar relational tasks in the OODB  
>> framework.
>>
>> I also think that Robert's lisp-as-query-language works well for  
>> the prevalence model when all objects are in memory, but I think  
>> it's less practical in, say, BDB where you are going to disk alot.   
>> However, it's a good discipline to consider - when does it makes  
>> sense to add new syntax/apis and when does it make sense to use  
>> lisp directly.
>>
>> You mentioned associations.  The best way to think about  
>> associations is that it is an easy way to maintain back pointers.   
>> For example, if a message object has a slot that contains a  
>> reference to a user, we may also want the user object to have an  
>> accessor that provides quick and efficient access to the collection  
>> of messages that point to it.  That's what associations are for.   
>> You could do this by declaring after methods on (setf (user  
>> message) value) that add the message to a pset sitting in a user  
>> instance slot, but that gets tedious.  As Leslie says, we're trying  
>> to make common cases simple and reasonably efficient.
>>
>> So the approach I'd like to see taken to designing the query  
>> framework is to capture the use cases and metaphors that people are  
>> really interested in and are encountering in real-world use and  
>> pick the largest subset that fits nicely into a clean, theoretical  
>> conceptual model.  There are already a good number (Leslie, Alex,  
>> etc) on the list that we could start with.
>>
>> For example, I often find myself wanting to filter a set of objects  
>> by more than one parameter (messages from user U that are high  
>> priority between 4/1/08 and 5/1/08).  What is the complexity of  
>> different approaches afforded by the existing Elephant  
>> implementation?
>>
>> In order of computational efficiency (I surmise):
>> 1. scan all messages and collect/operate on only those matching all  
>> criteria
>> 2. scan an index on messages instead of all messages; pick the one  
>> likely to yield the smallest subset
>> 3. intersection: scan two or more indexes for subsets represented  
>> as sequences of oids, instantiate, filter and operate on the  
>> objects represented by the intersection.
>> 4. create an index that orders objects by all three parameters and  
>> just walk the matching set.  Trade off space for time.
>>
>> Any others?
>>
>> The other consideration is the conceptual framework we want to use  
>> to approach the problem.  Procedural?  Constraint satisfaction?   
>> Logical form?  Graph matching?  There are some good examples of  
>> existing OODB systems in lisp out there (PLOB, AllegroStore/ 
>> AllegroCache, Statice, etc).  If you search the list archives, I  
>> think I've forwarded references in the past.
>>
>> I tend to lean towards a constraint satisfaction approach, as my  
>> sketch demonstrations.  "Operate on the set of objects that satisfy  
>> these constraints."  There are a bunch of practical issues.  Do we  
>> map query sets?  Do we cache them?  Do we represent them as lists?   
>> Are they lazily evaluated?  If we don't have a DSL, but allow  
>> arbitrary lisp expressions, then there isn't enough information to  
>> automatically select indexes, perform intersections, etc.
>>
>>
>> My other strong suggestion, besides starting by capturing the major  
>> use cases, is that we begin by implementing a procedural approach  
>> by implementing the building blocks for filter, sort, intersect,  
>> etc.  If we take the list of four filtering approaches above, we  
>> can start writing code that do these things and use them to  
>> implement some of the use cases.  The common building blocks and  
>> problems that we discover will inform the additions we'll want to  
>> the MOP, new implicit data structures like associations, the most  
>> convenient query syntax, etc.  Plus it will be useful in the  
>> meantime.  This fits into the classic lisp bottom-up DSL  
>> development model (well proselytized by Paul. Graham).
>>
>> Ian
>>
>>
>>
>> On May 9, 2008, at 6:02 PM, lists at infoway.net wrote:
>>
>>> Hello everyone,
>>>
>>> I apologize for being disconnected for so long. I had volunteered  
>>> to help in the query system and should have done more progress by  
>>> now. Unfortunately, the same as some (most or all) of you, putting  
>>> food on the table for my family has a higher priority and my  
>>> current job has demanded 110% of my time lately.
>>>
>>> Enough excuses! I have been passively reading several of your  
>>> email threads. I am convinced that a query system will bring a lot  
>>> of value to Elephant. The question that still arises is whether or  
>>> not people want a SQL-like syntax or a Lisp-like syntax.
>>>
>>> As Ian has suggested, publicly and/or privately, we should start  
>>> designing the query system in a very basic form. The most critical  
>>> part would be query optimization, which I'd rather work on after  
>>> we have the basic query system in place. But there are a lot of  
>>> decisions to make before we get there and coming to a consensus of  
>>> how it should look and how it should work is of critical importance.
>>>
>>> From a simplistic point of view, a SQL-like syntax should allow  
>>> for the execution of the basic relational algebraic operations  
>>> (union, difference, cartesian product, projection, and selection).  
>>> For the most part, these would not be difficult to implement.  
>>> However, IMHO, there is an intrinsic "contradiction" in applying a  
>>> SQL-like syntax on top of Elephant.
>>>
>>> Assume you have the following Tables (relations) in a SQL world:
>>>
>>> Books (
>>> book_id,
>>> title,
>>> author
>>> )
>>>
>>> Publishers (
>>> publisher_id,
>>> name
>>> )
>>>
>>> BooksPublishers (
>>> book_id,
>>> publisher_id,
>>> year
>>> )
>>>
>>> Suppose you wanted to get the cartesian product of all the books  
>>> published in 2008, you could run a SQL query like:
>>>
>>> SELECT Books.*, Publishers.* FROM Books, Publishers,  
>>> BooksPublishers WHERE Books.book_id = BooksPublishers.book_id AND  
>>> Publishers.publisher_id = BooksPublishers.publisher_id AND  
>>> BooksPublishers.year = 2008
>>>
>>> The result will be a concatenation of all the columns from the  
>>> Books and Publishers tables. In a SQL-world, you would access  
>>> these results in a key-value pair type mode (e.g. Books.book_id =  
>>> 1, Books.title = "1984", etc). However, when you think in terms of  
>>> Elephant (at least my understanding of it), you're dealing with  
>>> objects and not key-value pairs from multiple tables. So, instead  
>>> of getting a concatenation of all the columns, you "should" be  
>>> getting just a list of Book objects (or Publisher objects) that  
>>> met your query criteria, such that when you iterate thru them, you  
>>> could "query" their Publishers (or the Books). So, if we had  
>>> something like (please keep in mind this is no suggestion to  
>>> syntax or correctness but just for illustrative purposes):
>>>
>>> (defpclass book ()
>>> ((title :accessor book-title :index t)
>>> (author :accessor book-author :index t)
>>> (published_copies :accessor book-copies :initform (make-pset))))
>>>
>>> (defpclass publisher ()
>>> ((name :accessor publisher-name :index t)))
>>>
>>> (defmethod add-published-copy ((bk book) (pb publisher) year)
>>> (insert-item '(pb year) (book-copies bk)))
>>>
>>> (defmethod map-published-copies (fn (bk book))
>>> (map-pset fn (book-copies bk)))
>>>
>>> (setq objs (select book :where ((map-published-copies (lambda  
>>> (item year) (= (second item) year)) $bk 2008)))))
>>>
>>> From then on, you could just iterate through the book objects in  
>>> the result set for their respective published copies. The problem  
>>> with this is that, ok, you get all the books that met your  
>>> criteria but if you then wanted to get a list of all the published  
>>> copies, you would need to apply the filter criteria again. The  
>>> reason I think it "should behave" this way is because Elephant  
>>> deals with sets of objects, and you use Lisp to navigate through  
>>> the object space, whereas in a SQL-world you are not dealing with  
>>> objects but with a result set that contains all the columns you  
>>> asked for. If we were to emulate the same behavior in the query  
>>> system, that would sort of defeat the purpose of Elephant. For  
>>> that matter, you might as well use some of the other libraries  
>>> (e.g. CL-SQL, cl-perec, cl-rdbms, etc).
>>>
>>> The above example is a very simple example. We haven't looked at  
>>> SORTING, LIMIT, OFFSET, etc. Things which will simply make this  
>>> whole dilemma more difficult.
>>>
>>> I haven't looked into Ian's association mechanism yet. Maybe the  
>>> query system could/should be an extension to that with some  
>>> specialized features to apply filter criteria instead (and  
>>> possibly evolve into something similar to Ruby's ActiveRecord). I  
>>> know the association mechanism is still being developed and I  
>>> haven't really seen anyone comment much on it other than what Ian  
>>> has mentioned. In one of Ian's comments, he said:
>>>
>>>   "A more general query language is probably the right solution  
>>> for this interface.  The query language would know about  
>>> associations, derived indices, etc and perform query planning via  
>>> introspection over the class objects."
>>>
>>> At the same time, Robert said on another thread:
>>>
>>>   "One might philosophically prefer SQL.  I personally vaster  
>>> prefer to work in a powerful programming language to accomplish  
>>> these things. Obviously, whether two classes that refer to each  
>>> other stand in a "parent-child" relationship or not depends  
>>> entirely on the circumstances.  I prefer to write simple functions  
>>> such as "delete-order" below, which both utilize and (in a sense)  
>>> expand the power of LISP applied to persistent objects."
>>>
>>> Leslie said on yet another thread:
>>>
>>>   "While I'm at it: OFFSET and LIMIT (a real limit which lets you  
>>> specify an arbitrary Lisp expression) are things we definitely  
>>> want to aim for in 1.0. They are not difficult to implement at  
>>> all, but they don't work with GET-INSTANCES-BY-* and, worse, MAP- 
>>> BTREE. This means everyone has to write their own version of these  
>>> functions that take appropriate arguments and move the cursor  
>>> around themselves instead of relying on a simple high-level API.
>>>
>>> I'd have implemented these extensions myself, but I thought it  
>>> better to wait for the integration of the query language to add it."
>>>
>>> And Alex said:
>>>
>>>   "I think main problem is not how it looks, but that query  
>>> language actually makes programming a lot easier."
>>>
>>> All those comments make sense. There seems to be a group agreement  
>>> that something is needed, but everyone has their own ideas of how  
>>> it should work. Both the query language and the associations are  
>>> still being developed, so if we get consensus no how these should  
>>> work, it may give a better direction to both feature sets. If  
>>> anyone has any comments or suggestion as to whether a query system  
>>> be of real interest/necessity and if so, which would be the  
>>> preferred query syntax and expected behavior, that would really  
>>> help.
>>>
>>> I'm willing to work on this in as much as possible with my limited  
>>> knowledge of Lisp and Elephant. However, given a clear direction  
>>> of where this should go, I will be able to focus better and learn  
>>> faster what I haven't learned so far.
>>>
>>> Again, your feedback is much appreciated. I'm hopeful to be able  
>>> to work more on this over the weekend, assuming I get some  
>>> feedback from you guys.
>>>
>>> Thanks
>>> Daniel
>>> _______________________________________________
>>> elephant-devel site list
>>> elephant-devel at common-lisp.net
>>> http://common-lisp.net/mailman/listinfo/elephant-devel
>>
>> _______________________________________________
>> elephant-devel site list
>> elephant-devel at common-lisp.net
>> http://common-lisp.net/mailman/listinfo/elephant-devel
>
> _______________________________________________
> elephant-devel site list
> elephant-devel at common-lisp.net
> http://common-lisp.net/mailman/listinfo/elephant-devel